Information Processing Method for Determining Weight of Each Feature in Subjective Hierarchical Clustering

1. A method executed in a computer comprising a processor and a storage section in order to determine a weight of each feature in subjective hierarchical clustering of content expressed as a combination of physical features, the method comprising steps of: the processor acquiring multiple pieces of content separated into a triplet comprising a set of three pieces together with label information as training data, the label information indicating a pair specified by a user as having a highest degree of similarity among three pieces of the triplet, and storing the label information into the storage section; the processor executing hierarchical clustering using a feature vector of each piece of content of training data and a weight of each feature and determining a hierarchical structure of the training data; and the processor referring to the label information stored in the storage section to provide an updated weight for each feature so that a degree of agreement between a pair combined first as being in same clusters among three contents of the triplet in the hierarchical structure and a pair indicated by the corresponding label information increases.

2. The method according to claim 1 , further comprising the processor calculating an error rate which is a rate of a number of triplets for which a result of the hierarchical clustering and the label information do not agree with each other to a total number of the triplets; wherein the hierarchical clustering is repeatedly performed for the training data using the updated weight of each feature; and the processor outputs the updated weight of each feature in response to the error rate converging to a predetermined value.

3. The method according to claim 2 , wherein a weight is attached to each of the triplets; the step of determining the hierarchical structure comprises a step of executing, for each feature, hierarchical clustering on the weighted triplets using only one feature; the step of calculating the error rate comprises a step of calculating the error rate of the hierarchical clustering for each feature in consideration of the weight of each triplet; and the step of providing an updated weight of each feature comprises a step of updating the weight of a feature having the lowest error rate using the lowest error rate, updating the weight of each of the triplets with the updated weight of the feature, and outputting the updated weight of each feature in response to one of the hierarchical clustering result not being improved any more or the update processing having been performed a number of times corresponding to a total number of features.

4. The method according to claim 1 , wherein a weight is attached to each of the triplets; the step of determining the hierarchical structure comprises a step of performing, for each feature, hierarchical clustering on the weighted triplets using only one feature; the method further comprises a step of the processor determining an error rate of the hierarchical clustering for each feature in consideration of the weight of each triplet; the step of providing an updated weight of each feature comprises a step of determining a temporary weight of a feature having a lowest error rate using the lowest error rate, and providing the updated weight of each of the triplets as the temporary weight; and the processor repeating the step of performing the hierarchical clustering and the step of providing an updated weight of each feature, and outputs, for each feature, a sum of temporary weights determined for the feature as the updated weight of the feature in response to the hierarchical clustering result not being improved any more or the update processing having been performed a number of times larger than a total number of features.

5. The method according to claim 1 , further comprising a step of the processor acquiring one or more levels l for which a weight of each feature is to be switched and an upper-limit number of clusters N l for each level l; the step of determining the hierarchical structure comprises a step of performing hierarchical clustering in a bottom-up method until the upper-limit number of clusters reaches N l−1 using the weight of each feature determined for a hierarchical layer l−1, which is one level lower than the level l for which the weight is to be switched; and the step of performing the hierarchical clustering and the step of providing an updated weight of each feature are performed for the number N l−1 clusters as new training data, and the step of providing an updated weight of each feature comprises a step of outputting a determined weight of each feature as the updated weight of each feature for the level l.